U.S. patent application number 17/555850 was filed with the patent office on 2022-09-22 for methods and systems for analyzing and predicting aeroelastic flutter on configurable aircraft.
The applicant listed for this patent is Government of the United States, as represented by the Secretary of the Air Force, Government of the United States, as represented by the Secretary of the Air Force. Invention is credited to Brandon Baker.
Application Number | 20220300672 17/555850 |
Document ID | / |
Family ID | 1000006122381 |
Filed Date | 2022-09-22 |
United States Patent
Application |
20220300672 |
Kind Code |
A1 |
Baker; Brandon |
September 22, 2022 |
Methods and Systems for Analyzing and Predicting Aeroelastic
Flutter on Configurable Aircraft
Abstract
Methods and systems for analyzing and predicting aeroelastic
flutter on configurable aircraft are disclosed herein. The method
may include the steps of: a) flying a known aircraft type above
ground, wherein the aircraft has a payload in a known
configuration; b) acquiring data from at least one sensor on the
aircraft while flying above ground; c) repeating steps a) and b)
with a different payload configuration; d) training a machine
learning predictive model for the aircraft type for aeroelastic
flutter using the collected data; and e) using the predictive model
to predict when aeroelastic flutter may occur on the aircraft type
when the aircraft has a payload in a new configuration for which
data from sensors was not previously collected with the aircraft in
flight.
Inventors: |
Baker; Brandon; (Kaysville,
UT) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Government of the United States, as represented by the Secretary of
the Air Force |
Wright-Patterson AFB |
OH |
US |
|
|
Family ID: |
1000006122381 |
Appl. No.: |
17/555850 |
Filed: |
December 20, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
63163074 |
Mar 19, 2021 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06F 30/15 20200101;
B64F 5/60 20170101; G06F 30/27 20200101 |
International
Class: |
G06F 30/15 20060101
G06F030/15; G06F 30/27 20060101 G06F030/27; B64F 5/60 20060101
B64F005/60 |
Goverment Interests
RIGHTS OF THE GOVERNMENT
[0002] The invention described herein may be manufactured and used
by or for the Government of the United States for all governmental
purposes without the payment of any royalty.
Claims
1. A method for analyzing and predicting aeroelastic flutter on
aircraft, said method comprising: a) flying a known aircraft type
above ground, wherein said aircraft has a payload in a known
configuration; b) acquiring data from at least one sensor on said
aircraft while flying above ground; c) repeating steps a) and b)
with a different payload configuration; d) training a machine
learning predictive model for said aircraft type for aeroelastic
flutter using said acquired data; e) using the predictive model to
predict when aeroelastic flutter may occur on said aircraft type
when said aircraft has a payload in a new configuration for which
data from sensors was not previously collected with the aircraft in
flight.
2. The method of claim 1 wherein step c) comprises repeating steps
a) and b) with a plurality of N different payloads with M different
possible locations on the aircraft for at least a plurality of N+M
unique test flights.
3. The method of claim 1 wherein step c) comprises repeating steps
a) and b) with a plurality of different payloads wherein one type
of payload is positioned at a given location on the aircraft for
each flight, and the payload is changed to a different type payload
on each subsequent flight for said given location.
4. The method of claim 1 wherein prior to training the predictive
model, the data are organized using at least one of the following
techniques: scaling the data to a common range for all data types;
categorically encoding non-numerical values; and filtering the data
to improve data integrity.
5. The method of claim 1 wherein a machine learning predictive
algorithm is selected prior to training the predictive model, and
said machine learning predictive algorithm comprises one of the
following types of algorithms: Linear Regression, Logistic
Regression, Naive Bayes, Linear Support Vector Machines (SVM),
K-Nearest Neighbor, Decision Tree, Kernel SVM, Gradient Boost
Trees, Random Forests, Stochastic Gradient, Neural Networks, and
Convolutional Networks.
6. The method of claim 1 wherein the data are split into a first
group for training the machine learning predictive model and a
second group for validating the machine learning predictive
model.
7. The method of claim 6 wherein the data are split into said
groups using the K-Fold Cross-Validation technique.
8. The method of claim 1 wherein the data used for training a
machine learning predictive model for a particular aircraft uses
payload configuration data to predict the airspeed at which a
flutter event will occur.
9. The method of claim 1 wherein the data used for training a
machine learning predictive model for a particular aircraft uses
payload configuration data and airspeed to predict the Boolean
value of whether or not a flutter event will occur.
10. The method of claim 1 wherein a plurality of data are collected
and archived in a data storage device or apparatus to be utilized
to train a predictive model at a later time.
11. A system for predicting and warning of the potential for
aeroelastic flutter on an aircraft, said system comprising: a
processing unit located in an aircraft, wherein said processing
unit has a predictive model code loaded thereon for predicting the
onset of aeroelastic flutter, wherein said processing unit is in
communication with at least one sensor on the aircraft and is
configured to receive data from said at least one sensor on the
aircraft; and a warning mechanism in communication with said
processing unit that provides a pilot with an indication of an
impending flutter condition.
Description
[0001] Pursuant to 37 C.F.R. .sctn. 1.78(a)(4), this application
claims the benefit of and priority to prior filed co-pending
Provisional Application Ser. No. 63/163,074 filed Mar. 19, 2021,
which is expressly incorporated herein by reference.
FIELD OF THE INVENTION
[0003] The present invention relates generally to methods and
systems for analyzing and predicting aeroelastic flutter on
aircraft and, more particularly, to methods and systems for
analyzing and predicting aeroelastic flutter on configurable
aircraft that are able to warn a stakeholder of the conditions that
may cause aeroelastic flutter.
BACKGROUND OF THE INVENTION
[0004] Many types of engineered structures, such as skyscrapers,
bridges, and aircraft airframes, are subject to vibrational
stresses which can be caused by aerodynamic forces due to wind, for
example, or due to the airspeed of an aircraft in flight. The
aerodynamic forces over such structures may cause an unstable
oscillatory aeroelastic deformation, or vibration, of the structure
referred to as flutter. Flutter may involve different types of
motion, or stress, such as bending or twisting, combinations of
which may be referred to as a mode (e.g., mode of deformation) or
vibrational mode.
[0005] Aircraft, particularly military aircraft can be configured
in many different manners due to weapons/ordnances, loads, guidance
systems, and tracking systems. This can change the configurations
of the profile of the wings and/or the body of the aircraft. The
different configurations can each have their own set of conditions
under which flutter may develop.
[0006] Prior attempts to predict whether aircraft will develop
aeroelastic flutter are described in the patent literature. Some
patent publications describe the use of wind tunnels to attempt to
determine the wind speeds at which flutter initially occurs.
However, it is very impractical to conduct such testing for
different aircraft, each having many possible configurations. In
addition, the initiation of flutter in an aircraft's wings may
cause damage to the wings.
[0007] Other attempts to predict whether aircraft will develop
aeroelastic flutter when in flight have often involved the use of
finite element analysis. It has been found that finite element
analysis produces results that are not always sufficiently
accurate.
[0008] The need for improved methods and systems of more accurately
predicting when an aircraft will develop aeroelastic flutter has
continued. In particular a need exists for methods and systems of
more accurately predicting when an aircraft will develop
aeroelastic flutter that are capable of utilizing actual flight
data from the aircraft, and for receiving as input all relevant
data from the aircraft that might play a part in predicting when an
aircraft will develop aeroelastic flutter.
SUMMARY OF THE INVENTION
[0009] While the invention will be described in connection with
certain embodiments, it will be understood that the invention is
not limited to these embodiments. To the contrary, this invention
includes all alternatives, modifications, and equivalents as may be
included within the spirit and scope of the present invention.
[0010] According to one embodiment of the present invention, a
method for analyzing and predicting aeroelastic flutter on an
aircraft is provided. The method comprises: [0011] a) flying a
known aircraft type above ground, wherein the aircraft has a
payload in a known configuration; [0012] b) acquiring data from at
least one sensor on the aircraft while flying above ground; [0013]
c) repeating steps a) and b) with a different payload
configuration; [0014] d) training a machine learning predictive
model for the aircraft type for aeroelastic flutter using the
acquired data; [0015] e) using the predictive model to predict when
aeroelastic flutter may occur on the aircraft type when the
aircraft has a payload in a new configuration for which data from
sensors was not previously collected with the aircraft in
flight.
[0016] As further described in the Detailed Description, if any of
the steps (such as steps (a) to (c) have previously been completed
such that data exists for the same, then the method may start with
step (d) above.
[0017] According to another embodiment of the present invention, a
system for predicting and warning of the potential for aeroelastic
flutter on an aircraft is provided. The system comprises:
[0018] a processing unit located in an aircraft, wherein the
processing unit has a predictive model code loaded thereon for
predicting the onset of aeroelastic flutter, wherein the processing
unit is in communication with at least one sensor on the aircraft
and is configured to receive data from at least one sensor on the
aircraft; and
[0019] a warning mechanism in communication with the processing
unit that provides a pilot with an indication of an impending
flutter condition.
[0020] Additional objects, advantages, and novel features of the
invention will be set forth in part in the description which
follows, and in part will become apparent to those skilled in the
art upon examination of the following or may be learned by practice
of the invention. The objects and advantages of the invention may
be realized and attained by means of the instrumentalities and
combinations particularly pointed out in the appended claims.
BRIEF DESCRIPTION OF THE DRAWINGS
[0021] The accompanying drawings, which are incorporated in and
constitute a part of this specification, illustrate embodiments of
the present invention and, together with a general description of
the invention given above, and the detailed description of the
embodiments given below, serve to explain the principles of the
present invention.
[0022] FIG. 1 is a perspective view of a military aircraft.
[0023] FIG. 2 is a flow diagram showing one embodiment of the
method.
[0024] It should be understood that the appended drawings are not
necessarily to scale, presenting a somewhat simplified
representation of various features illustrative of the basic
principles of the invention. The specific design features of the
sequence of operations as disclosed herein, including, for example,
specific dimensions, orientations, locations, and shapes of various
illustrated components, will be determined in part by the
particular intended application and use environment. Certain
features of the illustrated embodiments have been enlarged or
distorted relative to others to facilitate visualization and clear
understanding. In particular, thin features may be thickened, for
example, for clarity or illustration.
DETAILED DESCRIPTION OF THE INVENTION
[0025] The present invention relates generally to methods and
systems for analyzing and predicting aeroelastic flutter on
aircraft and, more particularly, to methods and systems for
analyzing and predicting aeroelastic flutter on configurable
aircraft that are able to warn a stakeholder of the conditions that
may cause aeroelastic flutter.
[0026] The term "aircraft" refers to a machine that can fly. For
the purposes of the present invention, the term aircraft refers to
a flying machine that may experience aeroelastic flutter. There are
perhaps thousands of aircraft types, which may include but are not
limited to: BAC (Jet Provost, Strikemaster, TSR-2), Boeing YB-9,
General Dynamics F-16 Fighting Falcon, Lockheed Martin F22 Raptor,
Lockheed Martin F-35 Lightning II, Lockheed XFV, McDonnell Douglas
F-15 E Strike Eagle, and Northrop F-5, to name a few.
[0027] To distinguish the many airplanes produced of the same type,
a tail number may be assigned. For example, General Dynamics may
have produced roughly 4,500 F-16 Fighting Falcons, so there may be
4,500 unique tail numbers for that type of aircraft.
[0028] The term "stakeholder", as used herein, refers to a person
interested in analyzing, predicting, preventing, and/or being
warned of aeroelastic flutter on an aircraft. This may include, but
is not limited to: pilots, scientists, engineers, and researchers.
The stakeholder may be at various locations including, but not
limited to: in the aircraft during flight, or on the ground in an
office or laboratory with computing equipment, or in a flight
simulation environment.
[0029] In one embodiment of the present invention, a method for
analyzing and predicting aeroelastic flutter on an aircraft is
provided. The method comprises several steps. One version of the
method is shown in the flow chart in FIG. 2.
[0030] First Step--Flying the Aircraft
[0031] A first step may comprise flying a known aircraft type above
the ground, wherein the aircraft has a payload in a known
configuration. The aircraft can comprise any suitable type of
aircraft that is potentially subject to aeroelastic flutter. The
aircraft can be a military aircraft or a civilian aircraft.
Typically, the aircraft will be a fixed wing military aircraft
since such aircraft may carry payloads in different configurations,
and may be operated at higher speeds, although home-built aircraft
have been known to flutter at speeds as low as 55 mph.
[0032] FIG. 1 shows one non-limiting embodiment of a configurable
military fighter aircraft 20. The fighter aircraft 20 comprises a
fuselage 22, a pair of wings 24, an engine 26, an air intake 28, a
cockpit having a bubble canopy 30, a vertical stabilizer 32 having
a rudder 34, and a pair of horizontal stabilizers 36. The wings 24
have an upper surface and an underside. One example of such a
fighter aircraft may have 11 locations for mounting weapons and
other mission equipment.
[0033] The payload (or load) can comprise various combinations of
items attached to the aircraft externally or internally that may
include, without limitation, nuclear missiles, air-launched cruise
missiles, rotary launchers, cruise missiles, anti-ship missiles,
heat-seeking air-to-air missiles (AAM), radar guided medium-range
AAM, air-to-ground missiles, rockets, guided bombs, glide bombs,
radios, internal rotary launchers, internal smart bombs, hypersonic
missiles, electronic countermeasures (ECM), navigation units,
targeting or weapons pods, fuel tanks, sensor or radar pods, or a
cannon. Alternate forms of payload may be known to those skilled in
the art. Each item of payload may be characterized by physical
properties including, without limitation, weight (kg), moment of
inertia, drag coefficient, torque applied at location of payload
attachment, length, height, or width. The payload may be joined to
any suitable portion of the aircraft. For example, the payload may
be joined to and project from the underside of the wings of the
aircraft (rack mounts). For a given military fighter aircraft there
may be dozens of positions to attach or join a load, and there are
many different types of loads that may be attached. This may be
result in a large number of permutations of possible loads and
locations of attachment of the same. It would be impractical to
attempt to test every possible permutation for when each
permutation might develop aeroelastic flutter.
[0034] The terms "attached" and "joined", as used herein, encompass
configurations in which an element is directly secured to another
element by affixing the element directly to the other element;
configurations in which the element is indirectly secured to the
other element by affixing the element to intermediate member(s)
which in turn are affixed to the other element; and configurations
in which one element is integral with another element, i.e., one
element is essentially part of the other element. The terms
"attached" and "joined" includes both those configurations in which
an element is temporarily joined to another element, or in which an
element is permanently joined to another element.
[0035] When it is said that the first step may comprise flying a
known aircraft type above the ground, it is understood that this
step may not need to be performed anew when a particular aircraft
with a payload in a known configuration has already been flown, and
data described herein already exists for that aircraft with that
particular payload configuration. In addition, the step of
obtaining data from flying the aircraft above the ground is
intended to be distinguishable from data generated from testing the
aircraft in a wind tunnel since the latter data may be inherently
less reliable than in-use conditions.
Second Step--Collecting Data
[0036] A second step may comprise acquiring data from at least one
sensor on the aircraft while the aircraft is flying above ground.
The term "acquiring", as used herein with respect to data, may be
used to describe collecting and/or recording data. The aircraft may
have a plurality of sensors. These may include, but are not limited
to: altimeters, air speed sensors, piezoelectric sensors, a GPS
sensor or recorder, and sensors that measure or observe weather
conditions (e.g., cloudy, atmospheric/barometric pressure, air
temperature, and air pressure). In certain embodiments, it may be
desirable to record/collect at least the air speed of the aircraft.
All other data may be optional. In other embodiments, it may also
be desirable to collect/record data on the flight of the aircraft.
These may include, but are not limited to: roll rate, pitch rate,
and yaw rate (all in degrees), and the G-force. The data will
typically be recorded/collected continuously at periodic intervals
during each flight of the aircraft. The different types of data may
be collected at different time intervals. For instance, some data
may be collected every ten seconds, and other data may be collected
every ten milliseconds.
[0037] When it is said that the second step may comprise acquiring
data from at least one sensor on the aircraft while the aircraft is
flying above ground, it is understood that this step may not need
to be performed anew when a particular aircraft with a payload in a
known configuration has already been flown, and data described
herein already exists for that aircraft with that particular
payload configuration.
Third Step--Repeating First and Second Steps
[0038] A third step may comprise repeating a) the first step and b)
the second step with a different payload configuration for the
particular aircraft. One exemplary embodiment of this step may
comprise flying an aircraft with a different payload configuration
that may be a different airplane (with a different tail number) but
of the same aircraft type as described in the paragraph above that
defines the term "aircraft".
[0039] The different payload configurations may include, but are
not limited to: providing the aircraft with different payloads
(e.g., different types of weapons/ordnance, or other types of
payloads); and joining the payload(s) to different locations on the
aircraft.
[0040] The third step may comprise repeating a) the first step and
b) the second step with a plurality of N different payloads with M
different possible locations on the aircraft for at least a
plurality of N+M unique test flights.
[0041] (If a mathematical function could be derived to describe the
flutter properties of different payload configurations, it would be
a non-linear, multi-variate function with N+M unknowns.
Theoretically, one would only need N+M unique equations to solve
the non-linear system. For example: 10 different payload types to
be placed on a total of 10 different locations would give 20 total
unknowns; yet the number of configurations (or permutations) for 10
payloads at 10 locations is 10 to the 10.sup.TH power. N+M is the
theoretical lower-bound for the required number of equations to
successfully derive a sufficiently simple mathematical
multi-variate equation. In practice, many more than the theoretical
lower-bound are employed to gain sufficient information to overcome
instrumentation noise, trivial cases, higher-order non-linear
terms, and factors that may distort the results.)
[0042] In some cases, the third step may comprise repeating steps
a) and b) with at least five different payload types in at least
five different locations on the aircraft all the way up to
repeating steps a) and b) with hundreds of different payload types
on a dozen or more different locations on the aircraft. In some
cases, the third step may comprise repeating steps a) and b) with a
plurality of different payloads wherein at least one location on
the aircraft is provided with a plurality of different payloads
(that is, one type of payload is positioned at the specified
location for each flight, but the payload is changed to a different
payload on each subsequent flight). For example, the third step may
comprise repeating steps a) and b) with at least five different
payloads on a single location on the aircraft.
[0043] When it is said that the third step may comprise repeating
a) the first step and b) the second step with a different payload
configuration, it is understood that this step may not need to be
performed anew when a particular aircraft has already been flown
with different payload configurations, and the data described
herein already exists for that aircraft with those particular
payload configurations. For example, data may be collected and
archived in a data storage device or apparatus (such as a computer
hard drive or server) to be utilized to train a predictive model at
a later time. If such data is already available, then the method
can start with step 4 below. In such a case, step 4 will be the
first step.
Fourth Step--Training a Machine Learning Model
[0044] The fourth step comprises training a machine learning
predictive model (or simply the "machine learning model") for the
aircraft type for aeroelastic flutter using the data described in
the previous steps.
[0045] It may be desirable to carry out several additional
preliminary steps prior to the fourth step. These include, but are
not limited to: organizing the data; selecting a machine learning
predictive model; and splitting the data into a first group for
training a machine learning model and a second group for validating
the machine learning model.
[0046] Organizing the Data.
[0047] The preliminary step of organizing the data may include, but
is not limited to organizing the data using at least one of the
following techniques: scaling the data to a common range for all
data types; categorically encoding non-numerical values; and
filtering data to improve data integrity.
[0048] Scaling the data refers to a process in which data collected
in different units, for example, is placed into common units. For
instance, in some cases longitude and latitude data may be
collected in feet, meters, and degrees. Scaling involves converting
all such data to common units or a common scale. Scaling may also
include normalization by dividing each value from a particular
sensor or other data source by the standard deviation of all the
data points of that sensor or source. This practice can align
values from each distinct data source with values of other data
sources, so one source is not weighted more heavily than
another.
[0049] Categorically encoding uses practices known to those skilled
in the art of machine learning which may comprise selecting a
"Categorical" input type that may affect the aeroelastic flutter
properties of the aircraft. Categorical data may be identified in
society with a number so that humans can identify them easily (such
as a numerical index), but the quantifiable value of number itself
means nothing. An example of such encoding may be a "tail number"
which is effectively a serial number for a particular aircraft.
Flight plans, post flight briefings, test reports, or other
documentation may indicate a tail number by numbered indices (1, 2,
3 . . . 10, for example) but the quantity "1" or "1.0000" means
nothing in a mathematical sense in this case. Likewise, tail number
10 has nothing to do with the actual numerical value "10,"
especially in relation to the other nine indices. It just happens
to arbitrarily be assigned the number ten which was an otherwise
unused numerical index. Thus, each possible option for that
category is expanded as a separate input variable, and then a
Boolean value may be used (a 1 or a zero to indicate which category
a particular tail number was flown for that particular data). When
tail number six is flown, all collected data will have zeros for
categories "Tail Number 1," "Tail Number 2," "Tail Number 3," {4,
5, 7, 8, 9, and 10} and a "1" for category "Tail Number 6."
[0050] Categorical encoding may also include using non-Boolean
values for convoluted input data such as payload type and payload
location. For example, if payload location were considered a
categorical input, one may create input categories for each
physical or measurable quantity for each payload location. In other
words, there may be categories such as "Payload Location 1 Weight,"
"Payload Location 2 Weight," "Payload Location 3 Weight," (and so
forth) as well as "Payload Location 1 Air Drag Coefficient,"
"Payload Location 2 Air Drag Coefficient," "Payload Location 3 Air
Drag Coefficient," (and so forth), as well as any other desired
quantity that can be associated with payload locations. The entries
in each of these categories may not be Boolean, but will be
meaningful numerical quantities.
[0051] Filtering the data may comprise: identifying and/or removing
outliers; filling data where a datum is missing or is an outlier;
and eliminating all data for a particular flight if the data
contain a missing datum. Identifying and removing outliers may
comprise filtering out "noise" or outliers in the data, if a
particular measurement is sensitive to noise. For example, if a GPS
sensor is known to have limited resolution and it incorrectly
indicates that an aircraft has not moved, when the aircraft is
known to have traveled ten miles, then this data will be filtered
out. Similarly, if another sensor has spurious spikes that are not
physically possible or even realistic, then these spikes are
filtered out as well. Filtering may take place prior to scaling so
that the scaled values are not distorted by outliers or missing
values.
[0052] Selecting a Machine Learning Predictive Model.
[0053] The preliminary step of selecting a machine learning
predictive model comprises selecting the machine learning
predictive algorithm from a number of different types of
algorithms. Various types of algorithms include, but are not
limited to: Linear Regression, Logistic Regression, Naive Bayes,
Linear Support Vector Machines (SVM), K-Nearest Neighbor, Decision
Tree, Kernel SVM, Gradient Boost Trees, Random Forests, Stochastic
Gradient, Neural Networks, Convolutional Networks, and other types
of algorithms described herein.
[0054] The process of selection of an effective machine learning
algorithm depends on several factors. Some selection criteria may
include, without limitation: size of training data (number of
observations) compared to the number of features; interpretability;
model accuracy; training time or available computing resources;
linearity (or complexity) of input data; cleanliness of input data;
and output data type.
[0055] If the number of observations is small compared to the
number of features, one may select Linear Regression, Naive Bayes,
or Linear Support Vector Machines (SVM). As a minimum, 10
observations per variable is required for using any of these
methods. Such solutions may lack robustness when facing the
challenge of predicting outcomes for 10 to the 10.sup.th power
permutations, and thus may be less suitable. Conversely, if the
number of observations is large compared to the number of features,
a K-Nearest Neighbor, Decision Tree, or Kernel SVM algorithm may be
suitable. K-Nearest Neighbor method may not respond well to
non-linearities in the system and Decision Trees have been known to
be unstable at times, so in such cases, Kernel SVMs can be
selected.
[0056] If one desires to understand intimately how each input value
affects the outcome (i.e. "interpretability"), a Linear Regression
("restrictive") algorithm may be selected. If understanding the
relationship between inputs and outputs is not as important as
accuracy (for an exploratory exercise, for example), a less
restrictive algorithm may be selected. Such techniques, however,
may not exhibit the accuracy desired to accurately predict flutter,
and thus, may be less suitable for determining whether flutter is
likely to occur.
[0057] If time to completion (needed in a few minutes) or computing
resources are limited, such as 200G Floating Point Operations
(FLOPS) or fewer, Naive Bayes or Linear and Logistic Regression
algorithms work well. If time to completion is not as important, or
sufficient computing resources are available, a more accurate
prediction may be realized using SVMs, Neural Networks, or Random
Forests. Accuracy is most likely a more important outcome than time
to completion, so SVMs, Neural Networks, or Random Forests may be
more useful than simpler methods.
[0058] Complex data sets tend to require more complex algorithms
such as Kernel SVM, Random Forest, and Neural Networks. Complexity
of an input data set may be estimated by fitting the data (inputs
to outputs) using a linear regression, logistic regression, or
linear SVM and check the residual ("error"). If the residual is too
high, the data may be complex enough to merit a more complex
algorithm.
[0059] If the observations have outliers (the data are not
particularly "clean") Random Forests may be the best suited
algorithm. If the output data comprise selecting between two binary
classifications (like determining if a particular email is spam or
not spam), Logistic Regression or SVMs may be selected. In the case
of aeroelastic flutter, Logistic Regression or SVM may be useful to
simply determine if flutter is likely to occur, or not (a Boolean
outcome). In some cases, it may be desirable to clean the data well
prior to employing the predictive model training exercise, in which
case Random Forests may not be as suitable as other algorithms.
[0060] If data are highly complex but a high number of observations
(10,000 or more) are available, Deep-Learning (Neural Network with
multi-layer perceptrons) may be selected. Gradient Boost Trees can
be beneficial if observation data contain many missing values or
has a high number of irrelevant observation data input types. In
some cases, certain types of the algorithms described herein may be
excluded.
[0061] Splitting the Data Into Two Groups.
[0062] The preliminary step of splitting the data into two groups
comprises splitting the data into a first group for training a
machine learning model and a second group for validating the
machine learning model.
[0063] The quantity of data that is split into each group can
involve placing between about 60% to 90% of the data into the first
group for training the machine learning model and the remainder of
the data into the second group for validating the machine learning
model. A comprehensive splitting technique such as K-Fold
Cross-Validation may be used for splitting. K-Fold splits the data
into K folds, then trains the data on K-1 folds and tests on the
one fold that was left out. It does this for all combinations and
averages the result of each instance.
[0064] Training and Validating the Machine Learning Model
[0065] The data used for training the machine learning predictive
model for a particular aircraft and the output of the training can
take the form of at least two different embodiments. In one case,
the data used for training the machine learning predictive model
for a particular aircraft uses payload configuration data to
predict the airspeed at which a flutter event will occur. (In other
words, the payload configuration is the only input, and the speed
at which a flutter event is likely to occur is the only output.) In
another case, the data used for training the machine learning
predictive model for a particular aircraft uses payload
configuration data and airspeed to predict the Boolean value of
whether or not a flutter event will occur. (In other words, the
configuration and airspeed are the inputs, and "Yes" or "No"
indications that a flutter event will likely occur are the
output.)
[0066] When using K-Fold Cross Validation, the statistical average
of all the different combinations of training exercises is the
result, which tends to optimize the balance of splitting the
training and validation sets. With K-Fold, all observations are
used for both training and validation, and each observation is used
once for validation. Using K=10 provides a nice balance between
computational complexity and validation accuracy.
[0067] The end result of the Training and Validation exercise is a
mathematical function that accepts the same list of input values as
those used in training--only the outcome is unknown.
Fifth Step--Using the Model to Predict Flutter
[0068] The fifth step comprises using the predictive model to
predict when aeroelastic flutter may occur on the aircraft type
when the aircraft has a payload in a new configuration for which
data from sensors was not previously collected with the aircraft in
flight. When one desires to predict a flutter event, one inputs
data into the mathematical predictive model using a computer
hardware or software combination. The mathematical model then
computes an estimate for the desired prediction value (lowest speed
at which a flutter event is likely to occur, or a Boolean flutter
event detection).
[0069] The method of the present invention may be provided in the
form of a system. It should be understood that several of the steps
of the methods described herein will be computer-implemented. Such
steps may include, but are not limited to: acquiring data; training
the machine learning predictive model; and using the predictive
model to predict when aeroelastic flutter may occur. Applying the
predictive model to predict a flutter event may occur in an office
or laboratory with computing equipment, a flight simulation
environment, or actually in-flight with real-time sensors feeding
the predictive model.
[0070] The present invention may also provide a system for warning
a pilot of the conditions that cause aeroelastic flutter. Most of
the computer power exercised in machine learning is in building the
predictive model. Once the predictive model has been built, the
model may be employed on an aircraft and used in flight. Running
the predictive model requires relatively few computing resources.
Aircraft sensor data captured in real-time in flight may be
transmitted to a processing unit with the predictive model code
loaded on it. The sensor data are fed into the predictive model
code as inputs and the output is the model's prediction, whether a
Boolean "warning" or a speed at which flutter is predicted to
occur. In a particular embodiment, a warning light, voice, icon on
a screen, numerical values indicating the predicted speed at which
flutter is likely to occur, or other method may provide a pilot an
indication of an impending flutter condition. If the aircraft's
speed, for example, is approaching the flutter speed for a
particular aircraft type, the pilot may be warned to reduce the
aircraft's speed.
[0071] The aeroelastic analysis methods and systems described
herein can provide a number of advantages. The methods have the
ability to determine when flutter may occur on a virtually
unlimited number of payload configurations, even if they have not
been tested in flight. The methods are more accurate than prior
attempts since they use actual flight data rather than conventional
finite element analysis or wind tunnel data. The methods are also
more practical than attempting to conduct wind tunnel tests on each
payload configuration. It should be understood, however, that these
advantages need not be required unless they are set forth in the
appended claims.
[0072] It should be understood that every maximum numerical
limitation given throughout this specification includes every lower
numerical limitation, as if such lower numerical limitations were
expressly written herein. Every minimum numerical limitation given
throughout this specification includes every higher numerical
limitation, as if such higher numerical limitations were expressly
written herein. Every numerical range given throughout this
specification includes every narrower numerical range that falls
within such broader numerical range, as if such narrower numerical
ranges were all expressly written herein.
[0073] While the present invention has been illustrated by a
description of one or more embodiments thereof and while these
embodiments have been described in considerable detail, they are
not intended to restrict or in any way limit the scope of the
appended claims to such detail. Additional advantages and
modifications will readily appear to those skilled in the art. The
invention in its broader aspects is therefore not limited to the
specific details, representative apparatus and method, and
illustrative examples shown and described. Accordingly, departures
may be made from such details without departing from the scope of
the general inventive concept.
* * * * *